AMD's Halo Station: 96 cores, no price, no ship date

Local AI inference hardware got a new headline machine at IFA 2026 on September 4. AMD revealed the Threadripper Halo Station, called it the most powerful workstation in the world, said it can run trillion-parameter models, and did not announce a price or a ship date. Both of those omissions are the story.

What actually happened

The configuration, per Tom's Hardware: 96 cores, dual liquid-cooled Instinct MI350P accelerators, aimed squarely at Nvidia's DGX Station. ServeTheHome lists the CPU as a Threadripper Pro 9995WX, up to 2TB of DDR5, and 144GB of HBM3E per MI350P — 288GB in the shown two-card build, with a path to four cards and 576GB. StorageReview puts it at 96 cores and 192 threads on eight-channel DDR5.

What AMD did not say: price, availability, OEM partners, operating system, or the software stack. ServeTheHome's spec table carries "TBC" in several rows. StorageReview's power figure and its trillion-parameter memory math are the outlet's own arithmetic, not AMD claims. Outlets have published street-price estimates north of six figures based on component pricing; those are estimates, and we are not going to repeat them as facts.

So: a real machine, publicly shown, with the two numbers that determine whether you can buy it left blank.

Why local AI inference hardware matters for your business

You are probably not buying this box. That is fine — it is still a useful signal, because it tells you the ceiling of the "run it yourself" option is rising fast, and the honest question underneath it is one you can answer today.

The question is not "should we buy hardware." It is "what does our inference actually cost per month, and what would we pay to make it not leave the building?" Most teams cannot answer the first half. Pull ninety days of API invoices, split them by workload, and you will usually find one or two jobs carrying most of the bill — batch classification, document extraction, an embedding refresh. Those are the workloads where owned hardware has ever made sense: steady, predictable, non-bursty, and often small enough for a model that fits on far less than 576GB.

Everything else — the spiky, customer-facing, latency-sensitive traffic — stays on an API, because you are buying elasticity, not FLOPS.

The other reason to run local is not cost at all. It is that the data never leaves. If you handle records that make a compliance conversation awkward, a quantized open model on a machine you own can be the cheapest legal answer even when the compute math looks worse. Price that separately from the token math and be honest about which one is driving the decision.

Announcements without prices are marketing. Your invoice is data. Start there.

Sources: Tom's Hardware, ServeTheHome, StorageReview.